/sync-gbrain code stage fails on existing sources (gbrain ≥0.42 --confirm-destructive guard); use reindex-code in place #1985

Summary

/sync-gbrain's code stage fails for any already-registered source on gbrain >= 0.42. Both --full and --code-only (and plain incremental) error with a destructive-delete guard. The orchestrator (gstack-gbrain-sync.ts) refreshes a code source by removing and re-adding it, but gbrain 0.42 now gates gbrain sources remove behind --confirm-destructive. The orchestrator doesn't pass that flag, so the remove is refused and the whole code stage errors.

The source is not deleted (the guard blocks the remove), and the memory + brain-sync stages still succeed — but the code index never gets refreshed through the orchestrator. The only working refresh path is calling gbrain reindex-code directly.

Environment

• gstack: 1.57.10.0
• gbrain: 0.42.40.0
• Engine: PGLite (local-stdio MCP), single-machine macOS (darwin)
• Source already registered (created by a prior /setup-gbrain / --full run)

Repro

1. /setup-gbrain → local PGLite, repo policy read-write (registers source gstack-code-<repo> with N pages). This first run works because the source doesn't exist yet (register = pure add).
2. Make any change (or none) and run /sync-gbrain --full or /sync-gbrain --code-only again.
3. Code stage errors; memory + brain-sync still pass.

Actual output

gstack-gbrain-sync (full):
  ERR   code         source registration failed: gbrain sources remove gstack-code-autoleadai-6b2cc40d failed:
⚠️  This will permanently delete: 768 pages, 6,786 chunks, 6,786 embeddings

To proceed, pass --confirm-destructive (or use soft-delete: gbrain sources archive gstack-code-autoleadai-6b2cc40d).
To preview without side effects: --dry-run
 (1.9s)
  OK    memory       gbrain import: 3 imported, 112 unchanged, 0 failed (3.7s)
  OK    brain-sync   curated artifacts pushed (2.1s)
  SKIP  dream        call-graph state unknown (doctor unavailable)
  3 ok, 1 error, 0 skipped

Identical ERR code for --code-only.

Root cause

The orchestrator's code stage re-registers an existing source via gbrain sources remove <id> + gbrain sources add, instead of refreshing in place. gbrain >= 0.42 added a --confirm-destructive guard on sources remove (it would drop all pages/chunks/embeddings). The orchestrator predates that guard and doesn't pass the flag, so re-registration of any existing source fails.

This is distinct from #1734 (autopilot destructive auto-recovery) — here it's the orchestrator's own remove+add registration path tripping gbrain's new safety guard, on a non-autopilot manual sync.

Impact

/sync-gbrain --full and --code-only can no longer refresh the code index for a repo that was already indexed once. Users on gbrain 0.42+ get a confusing "permanently delete" error on a routine re-sync. The index is safe (remove is blocked), but the documented refresh verb is broken.

Workaround (confirmed working)

Refresh in place, no remove:

gbrain reindex-code --source <id> --yes

For memory + artifacts only: /sync-gbrain --no-code (those stages are unaffected).

Suggested fix

In gstack-gbrain-sync.ts's code stage, for an existing source skip the remove+add and refresh in place with gbrain reindex-code --source <id> --yes (already the path --full advertises). Only sources add for a genuinely new source. If a true re-register is ever needed, prefer gbrain sources archive or pass --confirm-destructive explicitly with a clear log line — never an implicit destructive remove on a routine sync.